Background
The client is an AI company reliant on large-scale training data, with sources mixing self-collected, third-party-labeled, open datasets, and open-source corpora. Before financing due diligence, the investor required an opinion on training-data ownership and legality; the team then discovered multiple coverage gaps in the authorization chain.
Key Challenges
- Purpose gap: Some authorizations covered only “display/analysis,” not “machine-learning model training,” exceeding the granted scope.
- Open-dataset trap: Several open datasets’ licenses explicitly excluded commercial training use; using them in a commercial model created infringement exposure.
- Personal-information contamination: Un-de-identified personal information was mixed into training data, triggering PIPL notice/consent and controller-liability issues.
- Labeling ownership: Outsourced labeling deliverables had unclear copyright and data-use allocations.
Geason’s Approach
- Purpose review: Checked authorization scope per data class, distinguishing “display/analysis” from “model training,” supplementing or replacing out-of-scope data.
- Open/Public dataset screening: Built a dataset-license review checklist, excluding sources that explicitly excluded training use.
- PIPL handling: De-identified or excluded contaminated personal information, rebuilding notice/consent and processing relationships.
- Labeling contract redesign: Rewrote labeling-service agreements clarifying deliverable copyright and data-use authorization.
The Non-Obvious Judgment
The industry commonly frames training-data risk as “whether authorization was obtained.” Our view: the real high-frequency minefield is “whether the authorization scope covers AI training as a purpose”—large volumes of data are authorized only for display or internal analysis, not model training; and many open datasets’ licenses explicitly exclude commercial training. This means even “freely obtained” data may constitute infringement. Training-data ownership review should be front-loaded to the data-collection stage, not remediated at due diligence, or underlying-asset defects will directly depress valuation.
Key Outcomes
- Established a training-data authorization-purpose review mechanism; out-of-scope data identified, supplemented, or replaced.
- Open and open-source datasets screened; sources excluding training use excluded.
- Personal-information contamination handled via de-identification and relationship rebuild; labeling deliverable ownership clarified by contract.
Client names and commercial data are anonymized. The rules, regulatory procedures and deliverable types described are as they occurred, not simplified.